Professional Documents
Culture Documents
SPIEDigitalLibrary.org/conference-proceedings-of-spie
ABSTRACT
This contribution focuses on the study and comparison of different design approaches for designing phase-only
diffractive optical elements (PDOEs) for different possible applications in laser beam shaping. Especially, new results and
approaches, concerning the iterative Fourier transform a'gorithm, are analyzed, implemented, and compared. Namely,
various approaches within the iterative Fourier transform algorithm (IFTA) are analyzed for the case of phase-only
diffractive optical elementswith quantizedphase levels (eitherbinaryor multilevel structures). First, the general scheme of
the IFTA iterative approach with partialquantization is brieflypresentedand discussed. Then, the special assortment ofthe
general IFTA scheme is given with respect to quantization constraint strategies. Based on such a special classification, the
three practicallyinterestingapproaches are chosen, furtheranalyzed, and compared to each other. The performance ofthese
algorithms is compared in detail in terms ofthe signal-to-noise ratio characteristic developments with respectto the number
of iterations, for various input diffusive-type objects chosen. Also, the performance is documentedon the complex spectra
developments for typical computerreconstruction results. The advantages and drawbacks ofall approaches are discussed,
and a briefguide on the choice ofa particularapproach for typicaldesign tasks is given.Finally, the two ways ofamplitude
elimination within the design procedure are considered, namely the direct elimination and partial elimination of the
amplitudeofthe complexhologramfunction.
Keywords: digital diffractive optical element, laser beam shaping, iterative Fourier-transform algorithm, phase-only DOE,
binaryDOE, multilevel DOE, computer generated hologram, diffractive elementsynthesis.
1. INTRODUCTION
a
Recently, beam shaping elements in form of diffractive optical elements (DOEs) or computer-generated holograms
(CGHs) have progressed very rapidly, mainly because of the prominentdevelopmentof micro fabrication techniques14.
Since their origination5, what have also been dramatically improved are the computer design and encoding methods and
algorithms whichare nowbecoming more and more availablenot only to specialistsin the field ofdiffractiveoptics,but also
a
to a much widercommunity ofopticalengineers. Also, rapid evolution in computing technologies gives in this sense large
possibilities of complicated computerdesigns and simulations of diffraction processes, which were not technically possible
just a few years ago.
Thus, the application field ofsuch DOEs and CGHs has also dramatically expanded, rangingnowadays from "classical
shaping"opticaldiffractive elements to optical information processing and optical interconnects applications, to namejust a
few. What is very critical for all such sophisticated applications is the final reconstruction qualityofdesignedDOEs, as well
as the diffractionefficiency as high as possible. Depending on a particularapplication,typically, some kind of trade-off
between these output qualities is required. On the other hand, some fabrication restrictions still persist and thus the
a
quantization as constraint to eithermagnitude or phase ofthe signal transmission must often takeplace. Preferably, due to
energetic reasons, the phase-onlydiffractive optical elements (PDOEs) are ofprimaryinterest(digital or even binarymicro
a
structuring of DOEs relief), although design studies of amplitude DOEs have also appeared6'7. Hence, there is
simultaneously a strong need for properoptimization algorithms which design as best DOEsas possible underthe limitations
and constrains given.
*
Correspondence: Email: richter@troja.fjfi.cvut.cz; Telephone: +4202-2191 2285;Fax:+4202-84684818
Laser Beam Shaping III, Fred M. Dickey, Scott C. Holswade, David L. Shealy, 75
Editors, Proceedings of SPIE Vol. 4770 (2002) © 2002 SPIE · 0277-786X/02/$15.00
The main idea ofthe basic IFTA algorithm is wellknown6'8'9, and there is no needto go to any details on a general level
here. Since we are interestedin amplitude-only input objects, the phase freedom (i.e. a random initial phase distribution)
within the signal window, as well as the amplitude freedom outside the window are applied. In general, depending on
particular degrees of freedom chosen, an input signal,
representedwithin the signal window, is transformed from
the object plane to the hologram plane using the relevant
transform (e.g. Fourier transform when the far-field DOEs
. are ofinterest, as in this paper,or Fresneltransform).
The continuous complex spectrum obtained in the
hologram plane is often materialised by micro structuring
the fabrication process, which is often capable of
(n) during
producingonly structures with discretely valued functions of
transmittance. Hence, one ofthe necessary designtasks is to
transform complexhologramspectrum functionto a discrete
2
form (mostly to binary or -level phase-only form). This
discretization can be done sequentially through the iterative
process. Hence, the hologramspectrum is partiallyquantized
by the IFTA algorithm within an each iteration step. This
partially quantized signal is then transformed to the
reconstruction plane using the inverse Fourier transform.
Obviously, such partialquantization introduces some noisein
the reconstruction plane. This noise is removed by replacing
Fig.1 Schematic drawing of a typical complex-spectrum the of a. newly calculated the
histogram example for a 4-level . phase DOE the situation . amplitude
. . noisy
. signal within.
. . . . .
indicatessomepoint n somewherein the middle ofthe iteration. signal window with the desired amplitude of the original
. . .
process. Set of values 1, -1, i, and —i, represents the target signal. This modified signal is used as an input to the next
iteration step.
points E(n) is a typical surrounding, and e(n) is a number
The whole IFTA iterative process, as explained above,
(angle)that defines the surroundingunambiguously. can be mathematically describedin general terms withinthe
n-th iteration step as follows (see Fig. 2): g (k,l) is the
signal at the input of the n-th iteration loop in the object domain, G (m,n) is its Fourier transform (denoted Fl) in the
spectral(hologram) domain:
G (m,n)= FT[g (k,l)] (1)
(k,l) and (m,n) describethe coordinates in the object and spectral domains,respectively. After applying thequantization
operator Q atthe spectraldomain,the structure J-J (in,n) is generated:
H (m,n) = Q[G (m,n)] (2)
Globally, the quantization represents a transformation of the hologram function to the desired form while maintaining
reconstructed imagequalitywithin defined constraints. In fact, such a transformation is done by iterative expansion oflocal
\
600 593
10000
500
cz
2 2400
(I)
300
5808
200
2851
0 20
/
40 60 80 100
108
/ 20 40 60 80 100
iteration iteration
(a) (b)
iteration
(c)
Fig.3 A typical SNR development with respect to the iteration number (a) for the approach (I) with direct amplitude
elimination, for a 4-level phase DOE (a case ofthe followingdata is plotted: designmatrix=256, signalwindow= 32, total
number ofiterations N= 100 , 4-level phasedesign, signal window location—offaxis). The value of2851 represents the final
SNR obtainedafter N= 100 iterations; (b) for the approach (II) withdirectamplitudeelimination, and with arepetition ofthe
whole partial quantization (N1 20, N2 I ). The value of593 represents the final SNR obtained after N 100 iterations; (c)
for the approach (III) with direct amplitude elimination, and with a repetition of an each partially-quantizing step
(N1= 5, N2=20). The valueof2436represents the final SNR obtained after N= 100 iterations. In cases(b) and (c),the same
parameters as in the case (a) were applied.
where the function angle of the complex argument X is defined as angle(X)= ço ; X = XIe'' . Clearly, there are several
possibilities of choosing the design strategy now. On one side, s(n)can be kept just simply constant throughout all
iterations, equal to the whole spectral range of interest (as 2'r ) for a given set of quantization levels. Therefore, a full
e(n) =
{[}I
Here, L is the numberofphase levels, andthe functions [.•1and {...}' are defined as follows,respectively:
(7)
[a]+={[al] aZ (8)
where [...] denotes the classicalfunction ofthe integerpart ofthe argument, Z is thesetofintegernumbers, and
{a}
1 Il aEZ
(9)
=l{a} aEZ
wheresimilarly {...} denotesthe function ofthe fractional partofthe argument.
Apparently, the parameters N1 andN2 considered cannot be chosen completely arbitrarily, they are restricted with
respecttoN: N must be divisibleby the product N1N2. Thereby, it can be shown that depending on the actualvalues of
N and N2 , all possible cases of e(n)canbe divided into four
3rX
general groups, called here (0) — (III). Each particularrealization
3 then, in fact, represents a class of IFTA modification. The
question ofpracticalimportance is which ofthese approaches are
2.5
1 worth further evaluation or even application with the IFTA
general scheme. Although a lot of effort has been recently
2
cc devoted to particular approaches and their improvements, this
z •1
Fig.4 A typical SNR development with respect to the As is well known,these approachesare not of much use within
iteration number for the approach (I) in the case of no
the optimization IFTA scheme since they do not converge'6; as a
phase elimination (target continuous phase, P= 1, i.e. matter of fact, they represents a direct quantization (also called
only the operator A is applied — direct amplitude the multilevel phase hard clip method), without any partial
elimination). The same parameters as in Fig. 3 were
applied. optimization, and can be practically considered useful only for
large values of L. Hence, it will not be considered any further
This approach,denotedfurtheras the one-step approach (I), has been studied previouslyin differentcontents2124'26, and will
alsobe a subject ofcomparison. The approach (II) existsfor 1 < N1 < N A N2 = , i.e. I
(11)
performing the partialquantization itself, and outer,repeatingthe wholeinner ioop. Finally, the last approach(III) existsfor
1<N <N A1<N2 <N,i.e.
(III)1<N<NA1<N2<N
1
= s(n)=1 —-- (13)
L[N2J N1j
Fig.5 A development of the PDOE complex spectrumfor the approach (I) for the same design parameters as in figure 3. Each
subfigure represents thecomplex-spectrum histogram situation afterthe 20th iteration, i.e. states atthe iteration numbers 0, 20, 40, 60,
80, and 100 are depictedfromthe left tothe right.
This is, in fact, a very general scheme, whichallows also for the situation where N > N1N2(while 1 < N2 , in contrastto the
case (II) where always N2 = I ), i.e. the wholestructure ofiterations ( N1N2) is repeated severaltimesitself. As we intend to
=
furthercomparethe basic limitingcasesonly, the case of N N1N2 is considered below inmore detail.
Altogether, as is alsocommon, all the approaches can be designated as eitherthe one-stepapproaches (cases0 and I) or
thetwo-step approaches (casesII and III,with a repetitionofthe whole one-step procedure).
In the next part, the concentration is given to a comparison ofthethree above specifiedand in a sense special extreme
IFTA approachesthat are ofpracticalinterest: (I), (II), and (III). These are, more specifically: (I) the one-step approach with
the direct partial quantization — N1 = N A N2 = 1 , (II) the two-step approach with the repetition of the wholepartial
quantization—1 < N1 < N A N2 = 1 , and(III)the thirdtwo-stepapproach with therepetitionofan eachpartially quantizing
step — 1 < N1 < N A 1 < N2 < N . That is, in the case (I), a quantization process is performed only ones during the whole
design process, in each step applying continuously expanding surroundings of the pre-described quantization levels
(typicallyin a simpledefinedway,e.g. linearlyincreasing their area). In the secondcase (II), as the firstmodification oftwo-
step approaches presented herein, the same one-step quantization process as in the case (I) is used (the whole partial
quantization takes place within the inner loop), but this procedure is itself repeated within the outer cycle. Finally, in the
approach (III), as the second modification oftwo-step approaches, all particularlyexpandedsurroundings are themselves
repeatedwhile the whole quantization is performed only ones.
Of course, new degrees of algorithm control arise within these modifications, i.e. particular numbers of repetition
chosen, and their combinations, can be tuned to obtain specificfinal results(in our case in terms ofthe SNR). It is obvious
that, in general, increasing the total numberof iterations would improve the final PDOEperformance, providedthat other
proper convergence conditions are fulfilled (e.g. the right shapeof the scale factor withinthe R operator, etc.). However,
for the purposeofthe comparison withinthis paper,in order to obtain the comparable results,the total numberof iterations
N would be strictly kept constant within this section for all approaches. This means that all three processeshave been
primarily studied withthe same total numberofiterations.
Fig.6 A development ofthe PDOE complex spectrum for the approach (II) forthe same design parameters as in Fig. 3. First, third,
and fifth cycle out ofthe total number offive outer cycles (N /N1= 5) are shown in the first, second, and third row ofsubfigures,
respectively. Within each ofthese sequences, 4th 8th 12th 16th, and 20th iteration situationis plotted.
of the total number of iterations(N chosen to be 100 in Fig. 3); then, during the second halfof the process, it rapidly
decreases to its final value (SNR= 2851). To explain this behaviourwe must consider the influence of the method of
Fig.7The testing PDOEobject used forthe referencecalculationspresented: (a) input object, (b) simulated computer reconstruction
ofthe input object, and (c) designed PDOE microstructure. Parameters: design matrix: 256 x 256, signal window:32 x 32, total
numberofiterations: N= 100 4-level phasedesign,signalwindow location- offaxis.
,
smaller than the final SNR value in the first case (compare final valuesof SNR= 2851 and 593, for the approaches (I) and
(II), respectively).
A case of the third approach (III), where each partial expansion of surroundings of quantization levels is repeated
several times while the whole quantization is made only ones, is depicted in Fig. 3c) ( N = 100 = N1N2 = 5.20 , tangibly
SNR = 2436). Here, we have concentrated on the special and most importantcase within the frame of (III) only, i.e.
N =N1N2 . The remaining case of N > N1N2represents, in fact, already a more complicated combined case, with the
featuresintrinsic to the case (II), namelythe repetitionofthe whole quantization procedure, resulting in a overalltendency
of the behaviourof the outer-envelope shape of the SNR curve. In this sense, certainly, further sub-classification of the
approach (III), underthe additional assumption apart from the given above, can be contemplated. However, furtherstudies,
especially on the final SNR performance are needed inthis combined case.
For the case considered herein, N = N1N2 , it is clearly seen that the repetition of elementary expansions results in
abruptincrease ofSNR, similarlyto the approach (I), but each "newly enlarged" expansion atthe beginning ofthe new inner
sub-cycle causes, in turn, an abrupt SNR decrease which is in a new cycle again compensated by repeating the same
Fig.8 A development ofthe PDOE complex spectrum for the approach(I) with iterativeamplitude elimination, for the same design
parameters as in Fig. 3. Each subfigurerepresents the complex-spectrum histogram situation after the 20th iteration, i.e. states at the
iterationnumbers 0, 20, 40, 60, 80, and 100 aredepictedfromthe left tothe right.
the first approach (SNR= 2851). In Fig. 5, each subfigure represents the situation after the next 20th iteration before the
actual partial quantization takes place (direct amplitude elimination and the scale factor applied). Such a spectrum
development is clearlydistinguishable from a case without the proper scale factor, as was demonstrated in Refs. 21,23-24,
and is a good evidence of the convergence quality. Indeed, in this case the PDOE spectrum clearly shows a very good
convergence ofthe approach (I) ofthe IFTA algorithm since the partially quantized structure is not practicallydeteriorated
by replacementof the amplitudewithin the signal window. It should be noted that each such a spectrum gives an ideal
reconstruction ofthe input object, and the last figure in the sequence illustrates the situation just beforethe final quantization
is proceeded. It is clearly seen that the spectrum is almost entirely prepared for the last quantization, i.e. it possesses only
desired quantized phase values.
Next, Fig. 6 shows the spectradevelopments forthe approach (II), i.e. the approach with arepetitionofthe wholepartial
quantization. It is shown how the spectrum evolves within the three cyclesout offive: 1st, 3rd(middle), and (final).Within
each ofthese sequences, 4th, 8th 12th 16th and 20th iteration are presented. Whenlookingat the last histograms withineach
cycle, it is evidenthowthe spreading ofthe processed values is gradually diminished. Finally, for the approach (III), i.e. the
approach with a repetition of an each partially quantizingstep, we have opted for an explanation without a figure. As a
matter of fact, if the spectra were taken each after the 20th iteration, they would look very similar to the first case (see
Fig. 5). This is also clear whencomparing the SNRdevelopments (Fig. 3a) and b)); the SNR curve inthe 1stcase is, actually,
an envelope of oscillating curve in thecase (III). In order to distinguish a detailedstructuring within spectradevelopments,
one wouldneed to plot in detail the spectraprogress in an inner loop, but this is notnecessaryhere.
In Fig. 7, the testing PDOE input object used for all the reference calculations presented above is show (Fig. 7a)),
together with the simulated computer reconstruction of the input object (Fig. 7b)), and designed PDOE microstructure
(Fig. 7c)), with the parameters specified in the figure caption. Here, the design was performed by the approach (I), but
clearly, the reconstructions and designed structures look at such a scale practically the same when generated by the
approaches(II) and (III). More importantly, choosinga differentdiffusive-type input objectwould lead to qualitatively the
same results, as the three approaches are compared; although the actual absolute numerical values of SNR would be
different, their relative ratia would be approximately maintained. Moreover, as is well known, a theoreticaldifference of
SNR (of orders ofhundreds or thousands) in several units or even tens ofpercent is not criticalfrom the practicalpoint of
view ofa PDOE design.
To conclude, the main message of this comparison is the following: for an efficientiterative design using the IFTA
scheme, all the presentedapproaches can be used in principle, depending on a designer's choice. The best results are
obtainedusing the approach (I) (see Fig. 3a)) when N1 = N A N2 = I , i.e. the local surrounding increment is as smallest as
possible (for a given N ). As is seen, this turns out to be a very important issue. However, the approaches(II) and (III) can
be competingwith the case (1). That has been confirmed by many testing examples. Additionally, the approaches (II) and
(III) provide a PDOE designer with a larger variability, namely with a variabilityofthe outer/inner cycles (choices of N1
convergence or to bring a better adaptability ofthe algorithm. This is due to the fact that a direct removal ofthe amplitude
always represents a radical intervention to the optimisation procedure(as is seen in Fig. 3a)); such a direct elimination
stronglydecreases SNR at the beginning ofthe iterative processwhereour aim is to minimize the overalldisturbance within
the step. Therefore, this idea of iterative amplitude removal, as implemented2324, has been applied to evaluatethe first
approach convergence with respect to the number of iterations. A typical result is shown in Fig. 8 for the iterative
developmentofthe complexhistogram when the amplitude is removed iteratively (the same parameters as in Fig. 5 were
applied).As can be seen, a good convergence has also been reached. However, a more detailedcomparisonis needed, as
well as an analysis of the partial amplitude elimination approach utilization, and of possibilities of combination of both
amplitudeelimination approachesforthe final SNR improvements. This will be a subjectofthe follow-up publication.
4. CONCLUSIONS
In particular, we have studied in a detail the phase operatorP as a part ofthe operator Q ofpartialquantization. We have
introduced a specific classification ofthe IFTA iterative scheme approaches with respectto quantization constraint strategies
that is based on the relationsand structuring ofthe loops within the total number of iterations N . In this sense, we have
presented a comparison of the three different extreme approaches within the iterative Fourier transform algorithm for the
case ofmultilevel phase-only diffractive opticalelements. It has turnedout to be beneficiary to studythese three approaches
separately, not only due to understanding the important tendencies and to mutual comparison. The approach (I), the
approach with a repetition of the whole partial quantization (II), and the approach with a repetitionof an each partially-
quantizingstep (III) have been used for designing PDOEs reconstructing the same types ofinput diffusiveobjects, keeping
the total number of iterationsalways the same. The performance ofthese three algorithms was compared using the SNR
characteristic developments as well as the complex spectrum histogram representations of a PDOE, with respect to the
number of iterations. Confrontation of the designapproach (I) with the two-stepapproaches(II) and (III) have shown that
thebest SNR resultscanbe obtainedusingtheapproach(I), but all thepresentedapproaches can be used in principle. The
a
main advantage ofthe two-stepapproaches is that they offer larger variability within the design process. The aim ofthis
paper, however, was not to select the prime approach, but rather to show differentcapabilities ofthese particularextreme
cases and, based on a mutual comparison, to enable a designerdeeper insight onto the iterative IFTA scheme modifications
and possibilities. Hence, the advantages and drawbacks ofall approacheshavebeen discussed whichcould imply a choice of
9. ACKNOWLEDGEMENT
This work has been partly supported by the Ministry ofIndustry and Trade ofthe Czech RepublicProject FB-C3/58. Also,
partial supports from the GrantAgencyof Academy of Sciences ofthe Czech Republic(with contract No. S 2065014)and
from the GrantAgency ofthe CzechRepublic (project202/011D004) are greatlyacknowledged.
5. REFERENCES
1. W. J. Dallas, Computer-generated holograms, in B.R.Frieden, Ed., The computerin opticalresearch, Topics in applied
physics,Vol. 41, Springer-Verlag,Berlin, 1980.
2. H. P. Herzig,Ed., Micro-optics: elements, systems andapplications,Taylor& Francis, London, 1997.
3. F. Wyrowski, Ed., SPIE Proceedings1718 - Workshopon Digital Holography, 1992; EOS Topical MeetingDigest
Series 12 - Diffractive Optics, 1997;OSA Technical Digest10 - Diffractive Opticsand Micro-Optics, 1998.
4. 0. Bryngdahl and F. Wyrowski, "Digitalholography— computer generatedholograms,"in Progress in Optics XXVIII,
ElsevierPublishing Company, Amsterdam, 1990.
5. B. R. Brown and A.W.Lohmann, "Complex spatial filteringwith binarymasks,"AppliedOptics5, p. 967, 1966.
6. F. Wyrowski, "Designtheoryofdiffractive elementsin the paraxial domain," J. Opt. Soc. Am. A 10, p. 1553, 1993.
7. R. Piestun, R., B. Spektor, B., and J. Shamir, "On-axis binary-amplitude computer generated holograms," Optics
Communications 136, p. 85, 1997.
8. F. Fetthauer, C. Stroot, and 0. Bryngdahl, "On the quantization of hologramswith the iterative Fourier transform
algorithm," Opt. Communications 136, p. 7, 1997.
9. J. R. Fienap, "Iterative method applied to image reconstruction and to computer-generated holograms," Optical
Engineering19, p. 297, 1980.
10. 5. Kirkpatrick, C. D. GelattJr., P. Vecchi, "Optimization by simulated annealing," Science 220, p. 671, 1983.
1 1. M. A. Seldowitz, J. P. Allebach, and D. W. Sweeney, "Synthesis ofdigitalholograms by directbinarysearch," Applied
Optics26, p. 71, 1987.
12. R. Hauck and 0. Bryngdahl, "Computer generated holograms with pulse-density modulation," J. Opt. Soc. Am. A 1, p.
5, 1984.
I3. K. Heggarty and R. Chevalier, "Signal window minimum average error algorithm for computer generated holograms,"
J. Opt. Soc. Am. A 15, p. 625, 1998.
14. J. N. Mait, "Fourier array generators," in H. P. Herzig, Ed., Micro-optics: elements, systemsandapplications, Taylor
& Francis, London, 1997.
I5. L. Legeard, P. Réfrëgier, and P. Ambs, "Multicriteria optimality for iterative encoding of computer-generated
holograms," AppliedOptics36, p. 7444, 1997.
16. F. Wyrowski, F. and 0. Bryngdahl, 0., 1988, IterativeFourier-transform algorithm appliedto computerholography. I
Opt. Soc. Am.A, 5, 1058.
I7. P. Birch, R. Young, M. Farsari, C. Chatwin, and D. Budgett, "A comparison of the iterative Fourier transform and
evolutionary algorithms for the design of diffractive optical elements," Optics andLasers in Engineering33, p. 439,
2000.
18. D. Lack, T. Zeggel, and 0. Bryngdahl, "Multi-level image quantization with object-dependent spectral control," Opt.
Communications 175,p. 33, 2000.
19. T. Tuttass, T. Zeggel, and 0. Bryngdahl, "Iterativealgorithm for image quantization with signal adaptation," Opt.
Communications 158, p. 18, 1998.
20. 5. Trester,"ComputersimulatedFresnel holography, Eur. I. Phys.21, p. 317, 2000.
21. M. kereñ, I. Richter, and P. Fiala, "Convergence studies of the iterative Fourier transform algorithm," EOS Topical
MeetingDigestSeries30, p. 30, 2001 (EOS TopicalMeetingon Diffractive Optics,Budapest, Hungary, 2001).
22. M. kerefi, I. Richter, and P. Fiala, "Fresnel-domain computer generated holograms," EOS Topical Meeting Digest
Series30, p. 76, 2001 (EOS Topical Meeting on Diffractive Optics, Budapest, Hungary, 2001).